Electricity marketing work order reply generation method and system based on generative large model

Abstract

The invention relates to an electricity marketing work order reply generation method and system based on a generative large model, and the method comprises the steps: obtaining an electricity marketing work order in any business scene, and carrying out the retrieval matching of the electricity marketing work order and a pre-constructed work order reply template database, and obtaining a recalled work order reply template; inputting the recalled work order reply template, the power marketing work order and the second cue word into a large language model to generate a standard work order reply; wherein the construction step of the work order reply template database comprises the following steps: acquiring power marketing work order reply data under multiple business scenes, and performing high-frequency template extraction to obtain candidate work order reply templates under the business scenes; and splicing the candidate work order reply template, a predefined work order reply template and the first cue word, inputting the spliced candidate work order reply template, the predefined work order reply template and the first cue word into the large language model, generating a standard work order reply template, and constructing a work order reply template database. Compared with the prior art, the method has the advantages of improving the flexibility and efficiency of work order reply automation and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of power marketing work order processing, and in particular to a method and system for generating power marketing work order replies based on a generative large model. Background Art

[0002] In the power marketing and customer service sectors, work order processing systems face complex challenges across multiple business scenarios. Traditional work order processing methods rely primarily on manual experience to create response templates and rule-based matching methods, making it difficult to address diverse business needs such as repair report acceptance, customer service responses, work order scheduling, and equipment inspections. This manually-led model is not only inefficient, but also often results in inconsistent work order response quality due to the complexity of business scenarios and the specialized nature of industry terminology, making it difficult to ensure consistent and accurate service. With the rapid growth of business volume, the limitations of traditional methods in terms of parsing efficiency, generalization capabilities, and adaptability have become increasingly prominent.

[0003] In recent years, the rapid development of generative artificial intelligence technology has revolutionized the processing of power marketing work orders. Generative AI based on large language models can understand the business scenarios and terminology unique to the power industry, and through semantic analysis and context modeling, it significantly improves the accuracy and adaptability of work order responses. This technology can dynamically generate professional and standardized response content based on specific business scenarios, effectively solving the pain point that traditional rule-matching methods are difficult to cope with complex business needs. The introduction of generative AI has enabled the work order processing system to have stronger semantic understanding and contextual reasoning capabilities, significantly reducing the need for manual intervention and providing key technical support for the intelligent upgrade of power marketing services. For example, patent application CN119599683A discloses an intelligent work order management system and method based on prompt word optimization.

[0004] However, the current generative AI power marketing work order processing has problems such as poor work order automation flexibility, slow generation speed, and poor generated text performance. Therefore, it is necessary to improve the existing processing technology to solve these problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for generating power marketing work order replies based on a generative large model, which improves the automation level and processing efficiency of work order replies.

[0006] The purpose of the present invention can be achieved by the following technical solutions: A method for generating responses to power marketing work orders based on a generative large model, comprising: Obtain power marketing work orders in any business scenario, search and match them with the pre-built work order reply template database, and obtain the recalled work order reply template; The recalled work order reply template, the power marketing work order, and the second prompt word are spliced ​​together and input into a large language model to generate a standardized work order reply; The steps of constructing the work order reply template database include: Obtain power marketing work order response data in multiple business scenarios, extract high-frequency templates, and obtain candidate work order response templates for each business scenario; The candidate work order reply template, the predefined work order reply template and the first prompt word are spliced ​​together and input into a large language model to generate a standardized work order reply template, and a work order reply template database is constructed.

[0007] Furthermore, the business scenarios include repair report acceptance, customer service response, work order scheduling and equipment inspection.

[0008] Furthermore, the step of obtaining a recalled work order reply template includes: Based on the power marketing work order, a pre-trained model is used to extract multiple keywords to form a keyword set; Based on the keyword set and the work order reply template database, the BM25 algorithm is used to perform template matching to obtain multiple nearest neighbor work order reply templates as recalled work order reply templates.

[0009] Furthermore, the step of forming a keyword set includes: Regarding the power marketing work order Perform word segmentation and part-of-speech tagging to identify nouns and obtain a set of candidate words ; Input the segmented power marketing work order into the pre-trained model to obtain the vector representation of each token in the sentence. The expression is: , Where, is the vector representation of token, is the number of tokens, is the vector dimension, for d dimensional column vector, is the field of real numbers; Based on the vector representation of each token, we take the average row by row to get the global vector representation of the sentence, which is expressed as: , Where, is the global vector representation of the sentence, is the averaging operation; For the candidate word set Candidate words in , take the vector average of all its occurrence positions in the sentence as the candidate word The vector representation of is: , Where, Candidate word The vector representation of Candidate word The vector representation of the first and second tokens in the sentence; Based on the global vector representation and candidate words of the sentence Vector representation of , calculate the candidate words The correlation with the sentence, where the calculation expression of the correlation is: , Where, For relevance; Sort the correlations from high to low, and select the top-k candidate words as the keyword set .

[0010] Furthermore, the step of obtaining multiple nearest neighbor work order reply templates includes: Define the work order reply template database as , combined with the keyword set, calculate the BM25 weight of the template in the power marketing work order and work order reply template database, where the calculation expression of the BM25 weight is: , in: , Where, is the BM25 weight, It is a work order reply template length, A collection of keywords, For terms The inverse document frequency of It is a term In the ticket reply template The frequency in ,avgdl is the average length of templates in the work order reply template database, and is the adjustment parameter, The number of templates in the work order reply template database, Represents the number of templates containing the term; Sort the BM25 weights from high to low and select the top-K nearest neighbor ticket reply templates. The expression is: , Where, Template for replying to nearest neighbor tickets.

[0011] Furthermore, the standardized work order reply is expressed as: , Where, To standardize work order replies, For large language models, A template for replying to a recalled work order. For power marketing work orders, is the second prompt word, Represents a splicing operation.

[0012] Furthermore, the LCS algorithm is used to extract the high-frequency template, and the specific steps include: 1) Set the power marketing work order reply data under the multi-business scenario as , and preprocessing is performed, where Reply data for the power marketing work order of the kth business scenario; 2) For the pre-processed power marketing work order reply data under each business scenario, the longest common subsequence (LCS) between two texts is calculated. At the same time, the dynamic programming method is used to record sub-problems to obtain the final longest common subsequence (LCS). The calculation expression of the longest common subsequence (LCS) is: , Where, Indicates the kth business scenario i Power marketing work order reply text, For the kth business scenario j Power marketing work order reply text, Represents a subsequence length; The operational expression of the dynamic programming method is: , Where, Indicates the i Power marketing work order reply text The first m length substring and the j Power marketing work order reply text The longest common subsequence of the first n length substrings, express The mth character of express The nth character of ; 3) Iterate step 2) to obtain the common subsequence set for each business scenario, expressed as: , Where, is the common subsequence set under the k-th business scenario; 4) Set the threshold , filtering out the low-frequency common subsequence set in each type of business scenario, obtaining the high-frequency common subsequence set in each type of business scenario, forming a high-frequency template and serving as the candidate work order reply template, wherein the high-frequency common subsequence set in each type of business scenario is expressed as: , formula, is the set of high-frequency common subsequences in the k-th business scenario, for The common subsequence in for exist The number of times it appears in for The number of.

[0013] Furthermore, the pre-processing step includes: The power marketing work order reply data in the multiple business scenarios are subjected to text cleaning, data desensitization and tag replacement processing, wherein special characters and garbled characters in the text are removed in the text cleaning processing, sensitive information in the text is marked as specific placeholders in the data desensitization processing, and regular expressions are used in the tag replacement processing to identify phrases containing time, place, and numbers in the text, and the phrases are replaced with specific placeholders.

[0014] Furthermore, the standardized work order reply template is expressed as: , Where, This is the standard work order reply template for the kth business scenario. For large language models, This is the candidate work order reply template for the kth business scenario. A predefined work order reply template for the kth business scenario. is the first prompt word in the k-th business scenario, Represents a splicing operation.

[0015] The present invention also provides a power marketing work order reply generation system based on a generative large model, comprising: Retrieval and matching module: used to obtain power marketing work orders in any business scenario, search and match them with the pre-built work order reply template database, and obtain the recalled work order reply template; Work order reply module: used to splice the recalled work order reply template, the power marketing work order and the second prompt word, input them into the large language model, and generate a standardized work order reply; Wherein, the search and matching module includes: High-frequency template extraction unit: used to obtain power marketing work order response data in multiple business scenarios, perform high-frequency template extraction, and obtain candidate work order response templates for each business scenario; Database construction unit: used to splice the candidate work order reply template, the predefined work order reply template and the first prompt word, input the large language model, generate a standardized work order reply template, and build a work order reply template database.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] (1) The present invention integrates knowledge from multiple business scenarios and guides the optimization and generation mechanism of reply templates through prompt words and a large language model, significantly improving the automation level and processing efficiency of power marketing work order replies.

[0018] (2) The present invention adopts the BM25 algorithm for template matching. By independently calculating the BM25 weight, it is not only applicable to templates containing multiple keywords, but also can adjust the influence of text length on the weight by adjusting parameters, avoiding the situation where the similarity of long texts is artificially high due to the large number of words. It is also friendly to long text processing and has the characteristics of strong interpretability and sensitivity to keywords, thereby improving the accuracy and efficiency of template matching.

[0019] (3) The BM25 algorithm of the present invention can efficiently process sparse data and highlight the importance of keywords through inverse document frequency (IDF). Even if the number of occurrences of keywords in the text is small, as long as the IDF is high enough, a high matching score can still be obtained, thereby further improving the accuracy of matching.

[0020] (4) The present invention adopts the LCS algorithm to extract high-frequency templates. When the LCS algorithm solves the longest common subsequence, a dynamic programming method is introduced to deal with the situation where the longest common subsequence problem has overlapping subproblems, thereby avoiding repeated calculations, reducing time complexity, and thus improving computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the overall process of the method of the present invention; Figure 2 A schematic diagram of a process for generating a work order reply template of the present invention; Figure 3 A schematic diagram of the work order reply generation process of the present invention. DETAILED DESCRIPTION

[0022] Example 1

[0023] This example provides a method for generating responses to power marketing work orders based on a generative large model. Figure 1 As shown, the following steps are included: S1: Collect power marketing work order response data from multiple business scenarios, extract high-frequency templates, and obtain candidate work order response templates.

[0024] This step is the template construction phase, which uses high-frequency template extraction technology and the LCS algorithm to generate candidate work order reply templates from multiple business work order data. Specifically: In this embodiment, the power marketing work order response data for multiple business scenarios includes response data for business scenarios such as repair report acceptance, customer service reply, work order scheduling, and equipment inspection.

[0025] Combine Figure 2 As shown, the power marketing work order reply dataset for multiple business scenarios is defined as , where each Represents the set of work order replies for the kth business scenario. Each business scenario is processed as follows: 1. Text cleaning: remove special characters and garbled characters.

[0026] 2. Data desensitization: Mark sensitive information such as customer information and employee information as specific placeholders.

[0027] 3. Token replacement: Use regular expressions to identify phrases such as time, place, and number, and replace them with specific placeholders.

[0028] A collection of work order responses for each business scenario (in For example), perform the following operations: Calculate the longest common subsequence (LCS) between two texts: , in, Represents a subsequence length, Represents the reply text of a power marketing work order. Since the longest common subsequence problem has the property of overlapping subproblems, a dynamic programming method can be used: , in, express The mth character of express The nth character of express The first m length substring of The longest common subsequence of the first n length substrings of . By iteratively calculating the longest common subsequence between two texts, we can get the common subsequence set: , Setting thresholds Filter the low-frequency fragments and keep the high-frequency LCS part: , in, It is The number of is the set of high-frequency common subsequences in the k-th business scenario, for The common subsequence in for exist The number of times it appears in . Extract and obtain candidate work order reply templates .

[0029] S2: Concatenate the candidate work order reply template, the predefined work order reply template, and the first prompt word, input them into the large language model, obtain the standardized work order reply template, and construct the work order reply template database.

[0030] This step is the template optimization phase, which combines candidate work order reply templates, predefined work order reply templates, and manually written first prompt words into the Guangming Big Language Model to generate standardized work order reply templates and build a work order reply template database. Specifically: The extracted candidate work order reply template , and predefined ticket reply templates and manually written first prompt words Input Guangming language model to generate standardized template : , in, Represents a splicing operation, Represents the Guangming language model.

[0031] Get a standardized work order reply template After that, it is stored in the database DB to obtain the work order reply template database.

[0032] S3: Extract keywords from the input power marketing work order, search the work order reply template database, and obtain the recalled work order reply template.

[0033] This step is the retrieval and matching stage. The input power marketing work order is tagged with parts of speech, noun phrases are identified as candidate words, and the candidate words are passed through the RoBERTa-wwm model. The scores are calculated by attention to obtain a keyword set. The keyword set is sparsely searched based on the BM25 algorithm to obtain the most relevant recalled work order reply template. Specifically: combine Figure 3 As shown, the power marketing work order under the newly input business scenario is defined as , input it into the pre-trained model (such as RoBERTa-wwm), perform part-of-speech tagging on its word segmentation results, identify nouns and obtain a set of candidate words , get the vector representation of the token in the sentence: , in is the number of tokens, is the vector dimension.

[0034] right Take the average by row to get the global vector representation of the sentence: , Candidate words (For example, after word segmentation), take the average of the vectors of all its occurrences in the sentence as its vector representation: , Calculate the relevance of candidate word c to the sentence: , according to Sort from high to low and take the top-k candidate words as the keyword set .

[0035] Define the work order reply template database as , calculate the BM25 weight of each work order reply template and power marketing work order: , in It is a term In the template The frequency in Is a template length, avgdl is the average length of text in the database, and is the adjustment parameter, is a set of keywords. The IDF formula is as follows: , in, is the total number of texts (templates) in the ticket reply template database, Represents the term The number of templates.

[0036] Sort the calculated BM25 weights from high to low and return the top-K nearest neighbor ticket reply templates as the recalled ticket reply templates: , Where, Template for replying to a recalled work order.

[0037] S4: The recalled work order reply template, the input power marketing work order, and the second prompt word are concatenated and input into the large language model to obtain a standardized work order reply.

[0038] This step is the response generation phase, which combines the recalled work order response template, the input power marketing work order, and the manually written second prompt word, inputs the Guangming Big Language Model, and generates a standardized work order response through the autoregressive Transformer model with a Decoder-Only architecture. Specifically: Reply template for recalled work orders , New Power Marketing Work Order The manually written second prompt word is input into the Guangming language model to generate a standardized work order reply: , Where, To standardize work order replies, The second prompt word is manually written.

[0039] The above method integrates knowledge from multiple business scenarios such as repair report acceptance, customer service replies, and work order scheduling. Through the template optimization and generation mechanism guided by prompt words, it significantly improves the automation level and processing efficiency of power marketing work order replies.

[0040] Example 2

[0041] This example provides a system for generating responses to power marketing work orders based on a generative large model, including: Retrieval and matching module: used to obtain power marketing work orders in any business scenario, search and match them with the pre-built work order reply template database, and obtain the recalled work order reply template; Work order reply module: used to input the recalled work order reply template, power marketing work order and second prompt word into the large language model to generate a standardized work order reply; Wherein, the search and matching module includes: High-frequency template extraction unit: used to obtain power marketing work order response data in multiple business scenarios, perform high-frequency template extraction, and obtain candidate work order response templates for each business scenario; Database construction unit: used to splice the candidate work order reply template, the predefined work order reply template and the first prompt word, input the large language model, generate a standardized work order reply template, and build a work order reply template database.

[0042] The rest is the same as Example 1.

[0043] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0044] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk drives, CD-ROMs, optical storage devices, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention may be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0045] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0046] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0047] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0048] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0049] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for generating responses to power marketing work orders based on a generative large model, characterized in that: include: Obtain power marketing work orders in any business scenario, search and match them with the pre-built work order reply template database, and obtain the recalled work order reply template; The recalled work order reply template, the power marketing work order, and the second prompt word are spliced ​​together and input into a large language model to generate a standardized work order reply; The steps of constructing the work order reply template database include: Obtain power marketing work order response data in multiple business scenarios, extract high-frequency templates, and obtain candidate work order response templates for each business scenario; The candidate work order reply template, the predefined work order reply template and the first prompt word are spliced ​​together and input into a large language model to generate a standardized work order reply template, and a work order reply template database is constructed.

2. The method for generating a reply to a power marketing work order based on a generative large model according to claim 1, characterized in that: The business scenarios include repair report acceptance, customer service response, work order scheduling and equipment inspection.

3. The method for generating a reply to a power marketing work order based on a generative large model according to claim 1, characterized in that: The steps of obtaining the recalled work order reply template include: Based on the power marketing work order, a pre-trained model is used to extract multiple keywords to form a keyword set; Based on the keyword set and the work order reply template database, the BM25 algorithm is used to perform template matching to obtain multiple nearest neighbor work order reply templates as recalled work order reply templates.

4. The method for generating a reply to a power marketing work order based on a generative large model according to claim 3, characterized in that: The step of forming a keyword set includes: Regarding the power marketing work order Perform word segmentation and part-of-speech tagging to identify nouns and obtain a set of candidate words ; Input the segmented power marketing work order into the pre-trained model to obtain the vector representation of each token in the sentence. The expression is: ， Where, is the vector representation of token, is the number of tokens, is the vector dimension, for d dimensional column vector, is the field of real numbers; Based on the vector representation of each token, we take the average row by row to get the global vector representation of the sentence, which is expressed as: ， Where, is the global vector representation of the sentence, is the averaging operation; For the candidate word set Candidate words in , take the vector average of all its occurrence positions in the sentence as the candidate word The vector representation of is: ， Where, Candidate word The vector representation of Candidate word The vector representation of the first and second tokens in the sentence; Based on the global vector representation and candidate words of the sentence Vector representation of , calculate the candidate words The correlation with the sentence, where the calculation expression of the correlation is: ， Where, For relevance; Sort the correlations from high to low, and select the top-k candidate words as the keyword set .

5. The method for generating a reply to a power marketing work order based on a generative large model according to claim 3, characterized in that: The step of obtaining multiple nearest neighbor work order reply templates includes: Define the work order reply template database as , combined with the keyword set, calculate the BM25 weight of the template in the power marketing work order and work order reply template database, where the calculation expression of the BM25 weight is: ， in: ， Where, is the BM25 weight, It is a work order reply template length, A collection of keywords, For terms The inverse document frequency of It is a term In the ticket reply template The frequency in ,avgdl is the average length of templates in the work order reply template database, and is the adjustment parameter, The number of templates in the work order reply template database, Represents the number of templates containing the term; Sort the BM25 weights from high to low and select the top-K nearest neighbor ticket reply templates. The expression is: ， Where, Template for replying to nearest neighbor tickets.

6. The method for generating responses to power marketing work orders based on a generative large model according to claim 1, characterized in that: The standardized work order reply is expressed as: ， Where, To standardize work order replies, For large language models, A template for replying to a recalled work order. For power marketing work orders, is the second prompt word, Represents a splicing operation.

7. The method for generating responses to power marketing work orders based on a generative large model according to claim 1, characterized in that: The LCS algorithm is used to extract the high-frequency template, and the specific steps include: 1) Set the power marketing work order reply data under the multi-business scenario as , and preprocessing is performed, where Reply data for the power marketing work order of the kth business scenario; 2) For the pre-processed power marketing work order reply data under each business scenario, the longest common subsequence (LCS) between two texts is calculated. At the same time, the dynamic programming method is used to record sub-problems to obtain the final longest common subsequence (LCS). The calculation expression of the longest common subsequence (LCS) is: ， Where, Indicates the kth business scenario i Power marketing work order reply text, For the kth business scenario j Power marketing work order reply text, Represents a subsequence length; The operational expression of the dynamic programming method is: ， Where, Indicates the i Power marketing work order reply text The first m length substring and the j Power marketing work order reply text The longest common subsequence of the first n length substrings, express The mth character of express The nth character of ; 3) Iterate step 2) to obtain the common subsequence set for each business scenario, expressed as: ， Where, is the common subsequence set under the k-th business scenario; 4) Set the threshold , filtering out the low-frequency common subsequence set in each type of business scenario, obtaining the high-frequency common subsequence set in each type of business scenario, forming a high-frequency template and serving as the candidate work order reply template, wherein the high-frequency common subsequence set in each type of business scenario is expressed as: ， formula, is the set of high-frequency common subsequences in the k-th business scenario, for The common subsequence in for exist The number of times it appears in for The number of.

8. The method for generating a reply to a power marketing work order based on a generative large model according to claim 7, characterized in that: The pre-processing steps include: The power marketing work order reply data in the multiple business scenarios are subjected to text cleaning, data desensitization and tag replacement processing, wherein special characters and garbled characters in the text are removed in the text cleaning processing, sensitive information in the text is marked as specific placeholders in the data desensitization processing, and regular expressions are used in the tag replacement processing to identify phrases containing time, place, and numbers in the text, and the phrases are replaced with specific placeholders.

9. The method for generating responses to power marketing work orders based on a generative large model according to claim 1, characterized in that: The standardized work order reply template is expressed as: ， Where, This is the standard work order reply template for the kth business scenario. For large language models, This is the candidate work order reply template for the kth business scenario. A predefined work order reply template for the kth business scenario. is the first prompt word in the k-th business scenario, Represents a splicing operation.

10. A power marketing work order reply generation system based on a generative large model, characterized by: include: Retrieval and matching module: used to obtain power marketing work orders in any business scenario, search and match them with the pre-built work order reply template database, and obtain the recalled work order reply template; Work order reply module: used to splice the recalled work order reply template, the power marketing work order and the second prompt word, input them into the large language model, and generate a standardized work order reply; Wherein, the search and matching module includes: High-frequency template extraction unit: used to obtain power marketing work order response data in multiple business scenarios, perform high-frequency template extraction, and obtain candidate work order response templates for each business scenario; Database construction unit: used to splice the candidate work order reply template, the predefined work order reply template and the first prompt word, input the large language model, generate a standardized work order reply template, and build a work order reply template database.

Images